1 / 51

Development of ATLAS B-taggers based on BDT

Development of ATLAS B-taggers based on BDT. Hai-Jun Yang, Xuefei Li & Bing Zhou University of Michigan, Ann Arbor ATLAS B-tagging Meeting May 6, 2009. Outline. Introduction B-jet decays with ‘long’ life-time MC samples (ttbar, WH120, WH400) B-tagging discriminating variables

hilda-zamora
Download Presentation

Development of ATLAS B-taggers based on BDT

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Development of ATLAS B-taggers based on BDT Hai-Jun Yang, Xuefei Li & Bing Zhou University of Michigan, Ann Arbor ATLAS B-tagging Meeting May 6, 2009

  2. Outline • Introduction • B-jet decays with ‘long’ life-time • MC samples (ttbar, WH120, WH400) • B-tagging discriminating variables • BDT B-taggers (for light-jets, C-jets, t-jets) • Performance comparison of ATLAS B-taggers • Cross checks of ATLAS B-taggers • Further improvement by combining B-taggers • Summary H. Yang - BDT B-tagging

  3. Introduction Goal: To develop ATLAS B-taggersusing BDT and to evaluate B-tagging performance using various B-tagger classifications (Likelihood, ANN, BDT). Our starting point: using the same set of variables as an initial try and to confirm results documented in a previous ATLAS note Ref: J. Bastos, ATL-PHYS-PUB-2007-019 H. Yang, talk at ATLAS B-tagging meeting on 02/09/09 based on ATLAS MC V13 samples Building a set ofnew discriminating variables to improve the b-tagging performance. Test the b-tagging performance using V14 MC samples H. Yang - BDT B-tagging H. Yang - BDT B-tagging 3

  4. B-jet decays with ‘long’ life-time • Primary Vertex • Secondary Vertex • Impact Parameter • Decay Length ~mm • No. of Tracks in jet Sign(IP) > 0 Sign(IP) < 0 H. Yang - BDT B-tagging H. Yang - BDT B-tagging 4

  5. MC Data Samples Ttbar (DS105200) based on release V14. Using WH(120), WH(400 GeV) with Hbb, cc, uu for test Preselection cuts: Et (jet) > 15 GeV, |h|<2.5 H. Yang - BDT B-tagging H. Yang - BDT B-tagging 5

  6. Input variables for BDT b-tagging algorithm Performance of BDT algorithm depends on selecting a set of discriminating variables; using advanced training process. Existing variables for an initial try, 11/17 are selected: IP2D, IP3D, SV1: jet weights from IP and secondary vertices Softe: jet weight from soft electron based tagger jet-mass: mass of particles which participate in the vertex fit Efrac: ratio between the total energy of charged particles in the vertex and the total energy of all particles in the jet d0sig_max, z0sig_max: the largest transverse and longitudinal impact parameter significance of tracks in the jet ptTrk_max: the largest transverse momentum of tracks in the jet Nvertex_2track: Number of two-track vertices Ntrack: Number of tracks in the jet H. Yang - BDT B-tagging H. Yang - BDT B-tagging 6

  7. NEW variables for BDT b-tagging algorithm Building additional 27 new variables, 9 / 27 are selected: Ntrack_distance_150: number of tracks in the jet with distance between PV and track-jet cross point greater than 150 microns Ntrack_z0_100: number of tracks in the jet with longitudinal IP greater than 100 microns Ntrack_z0_05: number of tracks in the jet with longitudinal IP significance greater than 0.5 2d_dl: 2D decay length of the jet (mm) d0sig_avg: average of transverse IP significance from two leading tracks which have the largest d0sig d0_avg: average of transverse IP from two leading tracks d0_avg: average of longitudinal IP from two leading tracks Sumtrkpt_jetE: ratio between sum of track Pt w.r.t jet axis direction and jet energy SumEpt_jetE: ratio between sum of Electron Pt w.r.t jet axis direction and jet energy H. Yang - BDT B-tagging H. Yang - BDT B-tagging 7

  8. Jet Weight from 3D Impact Parameters B-Jets C-Jets L-Jets t-Jets H. Yang - BDT B-tagging H. Yang - BDT B-tagging 8

  9. Numberof 2-track vertices B-Jets C-Jets L-Jets t-Jets H. Yang - BDT B-tagging H. Yang - BDT B-tagging 9

  10. Numberof tracks in jet B-Jets C-Jets L-Jets t-Jets H. Yang - BDT B-tagging H. Yang - BDT B-tagging 10

  11. Numberof tracks in jet with longitudinal IP > 100 microns (new) B-Jets C-Jets L-Jets t-Jets H. Yang - BDT B-tagging H. Yang - BDT B-tagging 11

  12. Numberof tracks with distance from PVto track-jet cross point > 150microns (new) B-Jets C-Jets L-Jets t-Jets H. Yang - BDT B-tagging

  13. Average of impact parameter significance from two leading tracks in jet (new) B-Jets C-Jets t-Jets L-Jets H. Yang - BDT B-tagging H. Yang - BDT B-tagging 13

  14. 2D Decay Length (new) B-Jets (tBc ~ 0.45mm) C-Jets t-Jets L-Jets H. Yang - BDT B-tagging H. Yang - BDT B-tagging 14

  15. BDT B-taggers • BDT_bl: B Jets vs Light Jets • BDT_bc: B Jets vs C Jets • BDT_bt: B Jets vs t Jets H. Yang - BDT B-tagging

  16. Btaggers: IP3D+SV1 and BDT_bl H. Yang - BDT B-tagging H. Yang - BDT B-tagging 16

  17. ttbar SamplesLight jet rejection vs b-tagging efficiency • For 60% B-Jet Efficiency, BDT B-tagger has ~ 50% improvement over IP3D+SV1 Light-Jet Rejection (B-tagger) ------------------------------------------- Light-Jet Rejection (IP3D+SV1) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 17

  18. BDT Performance vs Variables Used in BDT H. Yang - BDT B-tagging H. Yang - BDT B-tagging 18

  19. WH120 SamplesLight jet rejection vs b-tagging efficiency • For 60% B-Jet Efficiency, BDT B-tagger has ~ 40% improvement over IP3D+SV1 Light-Jet Rejection (B-tagger) ------------------------------------------- Light-Jet Rejection (IP3D+SV1) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 19

  20. WH400 SamplesLight jet rejection vs b-tagging efficiency • For 60% B-Jet Efficiency, BDT B-tagger has ~ 35% improvement over IP3D+SV1 Light-Jet Rejection (BDT) ------------------------------------------- Light-Jet Rejection (IP3D+SV1) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 20

  21. Comparison of B-taggers (only show the best three) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 21

  22. BDT B-taggers • BDT_bl: B Jets vs Light Jets • BDT_bc: B Jets vs C Jets • BDT_bt: B Jets vs t Jets H. Yang - BDT B-tagging

  23. Btaggers: IP3D+SV1 and BDT_bc H. Yang - BDT B-tagging H. Yang - BDT B-tagging 23

  24. ttbar SamplesC jet rejection vs b-tagging efficiency • For 60% B-Jet Efficiency, BDT B-tagger has ~ 40% improvement over IP3D+SV1 C-Jet Rejection (B-tagger) ------------------------------------------- C-Jet Rejection (IP3D+SV1) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 24

  25. WH120 SamplesC jet rejection vs b-tagging efficiency • For 60% B-Jet Efficiency, BDT B-tagger has ~ 30% improvement over IP3D+SV1 C-Jet Rejection (B-tagger) ------------------------------------------- C-Jet Rejection (IP3D+SV1) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 25

  26. WH400 SamplesC jet rejection vs b-tagging efficiency • For 60% B-Jet Efficiency, BDT B-tagger has ~ 30% improvement over IP3D+SV1 C-Jet Rejection (B-tagger) ------------------------------------------- C-Jet Rejection (IP3D+SV1) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 26

  27. Comparison of B-taggers (only show the best three) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 27

  28. BDT B-taggers • BDT_bl: B Jets vs Light Jets • BDT_bc: B Jets vs C Jets • BDT_bt: B Jets vs t Jets H. Yang - BDT B-tagging

  29. Btaggers: IP3D+SV1 and BDT_bt H. Yang - BDT B-tagging H. Yang - BDT B-tagging 29

  30. ttbar Samplest-jet rejection vs b-tagging efficiency • For 60% B-Jet Efficiency, BDT B-tagger has >~500% improvement over IP3D+SV1 t-Jet Rejection (B-tagger) ------------------------------------------- t-Jet Rejection (IP3D+SV1) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 30

  31. Comparison of B-taggers (only show the best three) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 31

  32. Cross checks of ATLAS Btaggers: Overlapped B-jets IP3D+SV1 BDT JetFitterCombNN • Apply cuts for IP3D+SV1, BDT and JetFitterCombNN with 60% of B jet efficiency, respectively. • Then calculate the overlapped B-jets passed these cuts. Overlapped efficiency = (A.and.B) / (A.or.B) H. Yang - BDT B-tagging

  33. Btaggers: Overlapped Light-jets IP3D+SV1 BDT • Apply cuts for IP3D+SV1, BDT and JetFitterCombNN with 60% of B jet efficiency, respectively. • Then calculate the overlapped light-jets passed these cuts. Overlapped efficiency = (A.and.B) / (A.or.B) JetFitterCombNN H. Yang - BDT B-tagging

  34. BDT B-tagger:Further Improvement  3 vars: IP2D, IP3D and SV1  11 vars: 11 existing variables  20 vars: 11 exisiting variables plus 9 new variables  22 vars: 20 vars plus IP3D+SV1 and JetFitterCombNN H. Yang - BDT B-tagging H. Yang - BDT B-tagging 34

  35. JetFitterCombNN and BDT_bl H. Yang - BDT B-tagging

  36. ttbar SamplesLight jet rejection vs b-tagging efficiency • For 60% B-Jet Efficiency, BDT B-tagger has ~ 120% improvement over IP3D+SV1 Light-Jet Rejection (B-tagger) ------------------------------------------- Light-Jet Rejection (IP3D+SV1) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 36

  37. Comparison of B-taggers (only show the best three) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 37

  38. BDT_bl vs IP3D+SV1(light jet rejection vs h) H. Yang - BDT B-tagging

  39. BDT_bl vs IP3D+SV1(light jet rejection vs PT) H. Yang - BDT B-tagging

  40. BDT_bl vs IP3D+SV1(B-Jet Efficiency vs h) H. Yang - BDT B-tagging

  41. BDT_bl vs IP3D+SV1(B-Jet Efficiency vs PT) H. Yang - BDT B-tagging

  42. Summary and Future Plan • B-tagger based on BDT technique has been developed and successfully implemented in offline Athena AnalysisExamples package. • By building 9 new discriminating variables, BDT B-tagger improves light jet rejection by a factor of ~1.5 compared to ATLAS default B-tagger (IP3D+SV1). • By combining JetFitterCombNN, the BDT B-tagger improves light jet rejection by a factor of ~ 2.2 compared to ATLAS default B-tagger (IP3D+SV1). • We are preparing a note and planning to implement BDT B-taggers in the ATLAS official Jet-Tagpackage. H. Yang - BDT B-tagging H. Yang - BDT B-tagging 42

  43. Backup Slides H. Yang - BDT B-tagging

  44. H. Yang - BDT B-tagging

  45. Eta and Pt of Jets H. Yang - BDT B-tagging H. Yang - BDT B-tagging 45

  46. List of Existing Variables IP2D – jet weight from transverse impact parameters IP3D – jet weight from 3D impact parameters SV1 – jet weigh from secondary vertices IP3D+SV1 H. Yang - BDT B-tagging H. Yang - BDT B-tagging 46

  47. More Existing Variables Softe – soft electron based tagger Mass – mass of particles which participate in vertex fit Efrc – energy ratio between particles in vertex and in jet N2t – No. of 2-track vertices H. Yang - BDT B-tagging H. Yang - BDT B-tagging 47

  48. More Existing Variables Largest IP2D – largest transverse IP significance of tracks in the jet Largest IP3D – largest longitudinal IP significance of tracks in the jet Largest Pt – largest transverse momentum of tracks in the jet Ntrk(jet) – track multiplicity in jet H. Yang - BDT B-tagging H. Yang - BDT B-tagging 48

  49. New Variables Ntrk(z0>100microns) Ntrk(z0sig>0.5) d0sig average of 2 leading tracks with maximum d0sig d0 value average of 2 leading tracks with maximum d0sig H. Yang - BDT B-tagging H. Yang - BDT B-tagging 49

  50. New Variables Z0 value average of 2 leading tracks with maximum z0sig Sum Pt of tracks which are related to jet direction / E(jet) Sum Pt of electron which is related to jet direction / E(jet) 2D decay length (mm) H. Yang - BDT B-tagging H. Yang - BDT B-tagging 50

More Related